PhoASR-whisper-small: Vietnamese Automatic Speech Recognition - A Revisit

We introduce PhoASR-whisper-small, a Vietnamese automatic speech recognition model obtained by fine-tuning the multilingual whisper-small on a 3000-hour dataset (PhoASR-3000h). The model directly produces clean transcripts with reliable timestamps, punctuation and capitalization, eliminating the need for separate post-processing tools. Please cite our PhoASR paper whenever PhoASR-whisper-small is used to produce published results or is incorporated into other software:

@inproceedings{phoasr,
    title = {{Vietnamese Automatic Speech Recognition: A Revisit}},
    author = "Vu, Thi and Nguyen, Linh The and Nguyen, Dat Quoc",
    booktitle = "Findings of the Association for Computational Linguistics: EACL 2026",
    year = "2026",
    pages = "6557--6568"
}

Model Inference with transformers

# Installation: pip install transformers==4.48.0
import torch
from transformers import pipeline

# Load the model
model_id = "Qualcomm-AI-Research/PhoASR-whisper-small"
device = "cuda" if torch.cuda.is_available() else "cpu"

# Create pipeline for automatic speech recognition with word-level timestamps
pipe = pipeline("automatic-speech-recognition", model=model_id, chunk_length_s=30, device=device, return_timestamps="word", generate_kwargs={"language": "vi"})

# Transcribe a single audio file
audio_path = "path-to-audio-sample.wav"
result = pipe(audio_path)

# Print full transcription
print(f"Full Text: {result['text']}")
# Print word-level timestamps
print(result['chunks'])

License/Terms of Use

This model is released under the BSD 3-Clause Clear license and the Qualcomm responsible AI license: https://www.qualcomm.com/site/responsible-ai-license

Uses

The model is intended for research and educational purposes.

Limitations and Bias

PhoASR-3000h is a variant of the PhoASR-3100h dataset (described in our paper above), excluding 100 hours of audio originally released under a CC-BY-NC-ND 4.0 license. Because only a relatively small portion of PhoASR-3000h contains Central regional accents, PhoASR-whisper-small may show reduced accuracy for Central Vietnamese speakers and other underrepresented dialects. It may also struggle to recognize emerging or rare terminology.